Benchmark

GPQA

Graduate-Level Google-Proof QA

Expert-level questions across physics, chemistry, biology that resist search-engine retrieval.

Models Tested

Gemini 2.5 Pro

Top Model

84%

Top Score

100 %

Max Possible

#	Model	GPQA Score	Provider	Composite
1	Gemini 2.5 Pro Google	84%	Google	97.5
2	Claude Opus 4 Anthropic	83.5%	Anthropic	97.8
3	o3 OpenAI	81.3%	OpenAI	97.1
4	o4-mini OpenAI	72.1%	OpenAI	91.5
5	DeepSeek R1 DeepSeek	71.5%	DeepSeek	92.3
6	Claude Sonnet 4 Anthropic	70.2%	Anthropic	90.3
7	Grok 3 xAI	68.5%	xAI	88.8
8	Claude 3.5 Sonnet Anthropic	65%	Anthropic	88.4
9	Llama 4 Maverick Meta	62.8%	Meta	87.5
10	Gemini 2.5 Flash Google	62.3%	Google	86.7
11	DeepSeek V3 DeepSeek	59.1%	DeepSeek	86.1
12	Qwen 3 235B Alibaba	58.3%	Alibaba	84.5
13	Gemini 2.0 Flash Google	55.4%	Google	80.5
14	GPT-4o OpenAI	53.6%	OpenAI	89.2
15	Llama 3.1 405B Meta	51.1%	Meta	82.1
16	Llama 3.3 70B Meta	48.7%	Meta	79.8
17	Claude 3.5 Haiku Anthropic	48.2%	Anthropic	79.2
18	Mistral Large Mistral AI	46.5%	Mistral AI	76.5
19	Qwen 2.5 72B Alibaba	45.6%	Alibaba	76
20	Grok 2 xAI	44.8%	xAI	74.8
21	Llama 3.1 70B Meta	42.3%	Meta	74.3
22	GPT-4o Mini OpenAI	41.1%	OpenAI	75.8
23	Phi-4 Microsoft	38.5%	Microsoft	64.5
24	Mixtral 8x22B Mistral AI	33.5%	Mistral AI	65.2
25	Command R+ Cohere	30.2%	Cohere	60.8

Source: Rein et al., 2023